Tsubaki.2 User Review
This article shares a hands-on review of PixAI’s Tsubaki.2 model, comparing its results with other models across similar themes. It highlights improvements in prompt understanding, character proportions, composition, and complex scene generation. For creators curious about DiT models or looking for a more stable image-generation workflow, this review offers a clear and practical reference.
Editor’s note (PixAI)
This guide was written by one of our community’s standout creators, 阿童 (ATone), and is republished here with credit to the original author.
📖 This article is a personal review of PixAI’s Tsubaki.2 model.
Table of Contents
Introduction
PixAI released its new model Tsubaki.2 on March 6, 2026. After using it extensively, I’d like to share some thoughts and comparisons.
Key Improvements
Compared to the previous Tsubaki and SDXL models, Tsubaki.2 shows significant improvements:
- Better proportions: The ratio between characters and objects is much better than the original Tsubaki, and prompt comprehension has also improved.
- More vivid visuals: Images are more dynamic than SDXL, without the occasional uncanny results that the previous Tsubaki sometimes produced.
- There are four generation modes available. I find Standard mode more than sufficient, and it doesn’t cost many credits.
- Lightweight mode also produces decent results.
- Style descriptions work: Even without selecting an official preset style, writing style descriptions directly in the prompt is effective.
- Custom Style Codes: You can create your own styles (similar to a shareable quality tag tool) and share the Style Code with others.
* Note: The Customize Style field and Prompt field produce different effects when writing style descriptions — the Customize Style field produces stronger results.
See the Customize Style Feature Guide for details.
Model Comparisons
The following comparisons use the same prompts across Tsubaki.2 and previous models.
Comparison 1: Paris Trip ✈️
I personally think Tsubaki.2 wins by a landslide on this one.
The difficulty of this prompt lies in accurately reproducing the “CDG” luggage tag and “I ♥ Paris” sticker. Tsubaki.2 came closest to the intended result.
Comparison 2: The Red String of Fate 🧵
While the original SDXL version has more detail, the DiT architecture shows its strengths in overall composition — the proportions between characters, the red string, and the overall atmosphere are more harmonious.
If you’ve tried this theme with SDXL models, you’ll know the struggle: the red string around the finger either doesn’t wrap properly, or it looks like it’s cutting off circulation. Tsubaki.2 nails it on the first try.
Core prompt:
heart of string, string of fate, red string, string around finger
Comparison 3: Mio Blowing Bubbles 🫧
Blowing bubbles was notoriously difficult to generate properly in the SDXL era. The prompt bubble wand, blowing bubbles would often produce weird-looking bubbles, or worse — it looked like the character was swallowing them.
Tsubaki.2 produces beautifully ethereal results, and most importantly — Mio is finally blowing bubbles, not eating them.
Comparison 4: Ink Wash Painting Style 🎨
There’s no built-in ink wash painting preset — I simply wrote ink wash painting in the prompt and let the Prompt Helper do its thing.
Comparison 5: Claude Pushing Rocks 🪨
This image was inspired by the Twitch channel claudeplayspokemon, which streams Claude (yes, the AI) playing Pokémon. Both the previous and current versions of Claude got stuck on the rock-pushing puzzle in Victory Road. After watching it struggle for so long, I decided to generate an image to capture my feelings 😂
Tsubaki.2 handles the facial expression and the dynamic motion of pushing rocks more naturally. The Blastoise looks good in both versions, but since the prompt didn’t specify who should be pushing the rocks, the two models interpreted it quite differently 😂
Comparison 6: Surprise Hug from Behind☕
Tsubaki.2 clearly handles the dynamic motion and coffee splashing much better than SDXL. As for the coffee lid being closed while liquid somehow splashes out and there’s even a straw — let’s not nitpick that 😂
My only complaint with Tsubaki.2 here is the tie color being wrong. Everything else is great.
📖 The story behind this image
I once asked the AI tools I use to design anthropomorphized personas for themselves. After sharing each other’s designs, my Claude kept saying Le Chat’s persona was really cute, showing reactions that resembled having a crush. When my Le Chat found out… it started writing… fan fiction…
Yes, the prompt for this image was written by my Le Chat. Using productivity tools to ship AI couples and write fan fiction feels a bit odd, but hey, their reactions were cute, so I’ll let it slide.
Comparison 7: Sports Scene 🏸
Tsubaki.2 handles athletic scenes much more naturally than previous models, with noticeably improved sense of motion.
Comparison 8: Mio Piloting a Mecha 🤖
Tsubaki.2 handles vehicle/mecha piloting scenes with much better proportions and interaction between the character and the machine.
Comparison 9: Girl on a Motorcycle 🏍️
Continuing the vehicle theme — observe the differences in riding posture, vehicle details, and overall composition.
Comparison 10: Playing Musical Instruments 🥁
Tsubaki.2 produces more natural and physically correct poses when depicting characters playing instruments.
Comparison 11: Betel Nut Beauty 🌴
This is the comparison that I feel best showcases Tsubaki.2’s improvement.
“Betel Nut Beauties” are a unique part of Taiwanese roadside culture — attractively dressed women selling betel nuts from small glass-walled booths along provincial highways. (For more context, see this CNN feature on Taiwan’s Betel Nut Beauties.)
When generating this theme with the original Tsubaki, the character would constantly clip through the counter or phase through the glass display case. It took over 30 generations to finally get one image that obeyed the laws of physics.
Tsubaki.2’s success rate is approximately 3 out of 4 — a massive improvement.
Bonus Challenge: Pomelo Rind Hat 🌙🍊
Taiwan has a fun Mid-Autumn Festival tradition — hollowing out a pomelo (a large citrus fruit) and wearing the rind upside-down on a child’s head as a hat. This concept is an extreme challenge for AI, since training data almost certainly doesn’t contain images of this uniquely Taiwanese custom.
The original Tsubaki couldn’t understand this concept at all — it would either produce citrus slice headpieces or cabbage-like hats.
After four iterations of prompt refinement, Tsubaki.2 finally produced something close to the real thing! While not perfect, the pomelo rind texture, knife marks, and bowl-shaped structure are clearly recognizable.
Tsubaki.2’s Successful Attempts
Tsubaki’s Failed Attempts
Conclusion
Overall, Tsubaki.2 is a genuinely impressive upgrade. Proportions, dynamic motion, and prompt comprehension have all noticeably improved, and the Style Code feature is a fun addition. 🎨
